rohansiva/A4D-dataset

A4D Dataset

Curated image dataset for A4D: Functional Latent Spaces for Affordance Reasoning (CoRL 2026).

• Project page: https://a4dance-reasoning.github.io/
• Paper (arXiv): https://arxiv.org/abs/2606.05533

Description

This is the curated publish_dataset split used for the A4D demos and qualitative results in the paper. Each image is paired with binary labels across 10 affordances, capturing task-relevant object functionalities (e.g., "movable", "graspable") rather than appearance-based object categories.

Stats

• Images: 552
• Object classes: 29 (inferred from filenames — e.g. banana, cup, pot, chair, drill, hammer, knife, sofa, table, door, bottle, ...)
• Affordances: 10 — containable, graspable, liftable, movable, openable, pushable, rollable, stackable, supportable, traversable
• Annotation rows: 5,096 (image, affordance) pairs — each image is labeled against 8–10 of the 10 affordances
• Label balance: 2,146 positive / 2,950 negative

Format

dataset.jsonl contains one row per (image, affordance) pair:

{"image": "banana_000.jpg", "affordance": "containable", "label": 0}

image refers to a file in this repo; label is 1 if the affordance applies to the object in the image, 0 otherwise.

Citation

If you use this dataset, please cite:

@inproceedings{siva2026objectsenablearefunctional,
            title={What Objects Enable, Not What They Are: Functional Latent Spaces for Affordance Reasoning},
            author={Rohan Siva and Neel P. Bhatt and Yunhao Yang and Seoyoung Lee and Nishant Gadde and Christian Ellis and Alvaro Velasquez and Zhangyang Wang and Ufuk Topcu},
            year={2026},
            booktitle={Proceedings of the Tenth Conference on Robot Learning},
            address={Austin, TX, USA},
            publisher={PMLR},
      }